GenMol: A Drug Discovery Generalist with Discrete Diffusion (2501.06158v2)

Abstract: Drug discovery is a complex process that involves multiple stages and tasks. However, existing molecular generative models can only tackle some of these tasks. We present Generalist Molecular generative model (GenMol), a versatile framework that uses only a single discrete diffusion model to handle diverse drug discovery scenarios. GenMol generates Sequential Attachment-based Fragment Embedding (SAFE) sequences through non-autoregressive bidirectional parallel decoding, thereby allowing the utilization of a molecular context that does not rely on the specific token ordering while having better sampling efficiency. GenMol uses fragments as basic building blocks for molecules and introduces fragment remasking, a strategy that optimizes molecules by regenerating masked fragments, enabling effective exploration of chemical space. We further propose molecular context guidance (MCG), a guidance method tailored for masked discrete diffusion of GenMol. GenMol significantly outperforms the previous GPT-based model in de novo generation and fragment-constrained generation, and achieves state-of-the-art performance in goal-directed hit generation and lead optimization. These results demonstrate that GenMol can tackle a wide range of drug discovery tasks, providing a unified and versatile approach for molecular design.

• The paper introduces a novel integration of discrete diffusion with the SAFE molecular representation to enhance molecule generation.
• The paper demonstrates state-of-the-art performance in hit generation and lead optimization using a fragment remasking strategy.
• The paper outlines a scalable, non-autoregressive framework that accelerates drug discovery by reducing trial-and-error processes.

An Analysis of GenMol: A Versatile Framework for Drug Discovery

The paper in focus introduces "GenMol," a novel generalist molecular generative model designed specifically to enhance various stages of drug discovery, including molecule generation, hit generation, and lead optimization. The authors employ a discrete diffusion framework in conjunction with the Sequential Attachment-based Fragment Embedding (SAFE) molecular representation to overcome limitations observed in existing molecular generative models. The significant advancements GenMol offers are noteworthy as it tackles a wide range of scenarios in the drug discovery pipeline, demonstrating a leap in performance over previous GPT-based models on the SAFE representation.

Key Contributions and Methodological Framework

GenMol stands out primarily due to its integration of discrete diffusion with the SAFE molecular representation. This combination leverages non-autoregressive bidirectional parallel decoding, which enhances computational efficiency and allows the model to fully utilize the molecular context that is agnostic to specific token ordering. Furthermore, the introduction of fragment remasking under the discrete diffusion process serves as a potent strategy for exploring the chemical space. This process involves optimizing molecules by replacing selected fragments with masks and subsequently regenerating them, a technique that facilitates the discovery of novel chemical structures.

Experimental Results and Implications

GenMol was put through rigorous testing across multiple tasks mimicking real-world drug discovery challenges, such as de novo generation, fragment-constrained generation, and goal-directed molecular optimization. Its performance notably exceeded that of existing molecular generative models, achieving state-of-the-art results particularly in goal-directed hit generation and lead optimization tasks. One of the more striking findings was GenMol's ability to significantly outperform multiple task-specific baseline models, suggesting its versatility and efficiency.

The outcomes from the fragment-constrained generation tasks highlighted GenMol's capability to generate high-quality molecules while maintaining diversity, a critical balance in chemical generation tasks. Moreover, the implementation of fragment remasking proved superior to traditional token remasking methods, underscoring the practical utility of fragment-level operations over atom- or bond-level manipulations for effective exploration of chemical spaces.

Theoretical and Practical Implications

Theoretically, GenMol's application of discrete diffusion models within the context of molecular sequence generation marks a novel approach in drug discovery that moves away from heuristic-based token ordering. The reliance on fragment-based exploration aligns more closely with chemical intuition, enhancing the model's ability to generate viable drug candidates.

Practically, GenMol's framework could significantly streamline the drug discovery process, offering a more unified tool for molecular design. This is critical for accelerating early-stage drug discovery, reducing time and costs associated with lengthy trial-and-error processes. As the framework does not require fine-tuning for different tasks, it offers substantial scalability and adaptability for various drug discovery scenarios.

Potential Future Directions

Future work could extend GenMol by incorporating more complex chemical data and exploring additional representations of molecular structures. Another potential development could involve integrating GenMol into a fully automated drug discovery pipeline, coupled with in-vitro or in-silico testing to further validate the generated compounds' efficacy and safety. Additionally, exploring transfer learning capabilities could allow GenMol to adapt quickly to emerging pathogens or other areas of urgent need in pharmaceuticals.

In summary, GenMol contributes a versatile and efficient model to the drug discovery arena. Its successful leveraging of discrete diffusion models with fragment remasking positions it not only as a practical tool for current challenges but also provides a foundation for significant advancements in molecular generation and optimization strategies.